Investigation On The Fast High Angular Resolution Diffusion Magnetic Resonance Imaging Algorithm

Diffusion Magnetic resonance imaging(d MRI)is currently the only non-invasive technology to detect the diffusion information of water molecules in living biological tissues.Water molecules in living tissues are limited by the influence of nerve,cell and surrounding tissue structure,and their diffusion displacement distribution can reflect the microstructure of tissues.In order to accurately estimate the diffusion displacement distribution of water molecules,high angular resolution diffusion imaging(HARDI)is proposed.HARDI doesn’t make any assumptions about the diffusion displacement distribution of water molecules,and directly uses more diffusion gradient direction signals to reconstruct the multi-fiber cross structure,which is an important means of nondestructive testing complex fiber structure.However,HARDI imaging requires the acquisition of diffusion weighted(DW)signals in many directions,which greatly increases the imaging time.In addition,the current HARDI reconstruction algorithm can’t estimate the orientation of small-angle cross fibers accurately.In order to solve the above problems,this paper designs two deep learning models to achieve fast HARDI imaging and accurate reconstruction of cross fibers.The specific research contents are as follows:(1)A direction multiplying network(DMNet)based on sampling is designed to generate multi-direction DW signals from few-direction DW signals to achieve fast HARDI imaging.The main structure of DMNet is composed of the multi-scale upsampling decoder module(MSUD)and the densely dimension-reducing residual fusion module(DDRRF),in which MSUD is composed of three up-sampling convolution paths of different scales,and rich directional features with multiple receptive fields are obtained through multi-channel fusion.The fused feature images are extracted and dimensionally reduced by DDRRF,and residual fusion is used to reconstruct multidirection DW signals.The model applies image spatial loss,wavelets loss and direction constraint correction loss for optimization training.By comparing the generated DWI and fiber reconstruction results with the actual multi-direction acquisition results,the experiment shows that the method is effective in generating multi-direction signals from few-direction signals and the acquisition time is reduced while the quality of fiber orientation reconstruction is guaranteed.(2)A multi-fiber reconstruction network(HARDI-Net)model is proposed for the reconstruction of voxel cross fibers.This method uses 1D-CNN to extract key features of multi-direction DW signals,mainly including multi-scale pooling feature extraction module and multi-path feature fusion module.In the process of down-sampling,multiscale pooling is used to reduce feature loss.At the same time,multi-scale features are fused for further feature extraction,and then multi-path features are fused and mapped to fiber orientations.The model uses fiber orientation bias loss and cross angle loss constraint network convergence.Comparing cross fibers reconstruction results of different angles between HARDI-Net and the reality,the angle errors are less than 5°.And compared with several existing reconstruction methods,the angle errors of HARDI-Net are far less than these methods.The results show that the proposed HARDI-Net can not only accurately reconstruct the orientation of large-angle cross fibers,but also has a good adaptability to small-angle cross fibers.